Research map / Vision-language and multimodal / Multimodal LLMs

Revisiting the Data Sampling in Multimodal Post-training from a Difficulty-Distinguish View

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes two difficulty metrics for multimodal samples (PISM via image masking sensitivity, CMAB via text/image attention balance) and finds GRPO-only on difficulty-stratified data beats SFT+GRPO for Qwen2.5-VL-7B.

Problem

Existing multimodal RL post-training lacks quantifiable difficulty metrics that account for cross-modal features for screening samples, and its SFT+GRPO pipelines may not be optimal for jointly improving perception and reasoning.

Approach

Two difficulty-aware sampling metrics: PISM measures the masking ratio at which a model's answers become unreliable under progressive random pixel masking; CMAB measures the geometric-mean ratio of image-to-text attention during response generation. Samples are binned into easy/medium/hard/unsolved, and GRPO-only training on mid+hard samples is compared with SFT+GRPO variants.

Key results

• GRPO-only on PISM mid+hard samples reached 68.3 on MathVista and 77.8 on OCRBench, beating SFT+GRPO variants on the reasoning/perception tables.
• CMAB mid+hard GRPO-only reached 69.0 on MathVista and 0.550 on MMMU (visual reasoning table) versus at most 68.1 for SFT+GRPO variants.
• GRPO-only(mid+hard) reached 69.716 on HallusionBench (PISM perception table), above the best SFT+GRPO variant at 68.980.

Surprising findings

• GRPO-only training on difficulty-stratified samples outperforms conventional SFT+GRPO pipelines, suggesting SFT may be unnecessary.
• Random-sample GRPO-only is competitive with or sometimes better than mid+hard on several benchmarks (e.g., MMVet 53.257 vs 48.257 in PISM reasoning table), so the selection benefit is not uniform.

Limitations

• Results are reported as single-run point differences with small margins (often under 1 point) and no variance estimates.
• Relies on a single base model (Qwen2.5-VL-7B) and an unnamed benchmark source for the training data.
• The thresholds (tau=0.1, lambda_hard=0.4, lambda_easy=0.7, attention ratio cut-offs) are set by hand without sensitivity analysis.

Experimental setup

Datasets: multimodal perception-reasoning benchmark from prior work (unnamed)

Benchmarks: MathVista MMVet OCRBench HallusionBench MMMU MMStar

Metrics: accuracy score

Baselines: GRPO-only(fullset) GRPO-only(random) GRPO-only(unsolved) SFT+GRPO hybrid variants

Builds on: GRPO DeepSeek-R1 Qwen2.5-VL

Closest papers by meaning

• Understanding the Role of Hallucination in Reinforcement Post-Training of Multimodal Reasoning Models (CVPR 2026 · similarity 0.90)
• Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models (ICLR 2026 · similarity 0.89)
• Reinforcing Multimodal Reasoning Against Visual Degradation (NeurIPS 2026 · similarity 0.89)
• AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.89)
• GThinker: Towards General Multimodal Reasoning via Cue-Guided Rethinking (CVPR 2026 · similarity 0.89)
• LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.89)